Here is the number you will not find in a Novo Nordisk press release: in US commercial insurance claims analyzed through 2023, only about 23 percent of patients starting a GLP-1 medication were still filling prescriptions 12 months later. The trials that launched this drug class ran 68 to 72 weeks. The real-world majority is gone before the trial would have even ended.
My thesis is this: the entire public conversation about GLP-1 drugs has been built around trial-condition efficacy numbers applied to a real-world population that does not behave like trial participants. That mismatch is not a minor footnote. It is the central fact about how this drug class actually functions as a public health intervention — and it is the fact that gets systematically omitted from the announcement, the headline, and the social post.
Let me go through the data in layers, because different sources tell slightly different versions of the same story, and each version adds something the others miss.
What the claims data says
The most-cited real-world persistence analysis comes from a 2022 paper in Obesity by Ganguly and colleagues, which examined over 3,000 US patients initiating GLP-1 receptor agonists for obesity (not diabetes) through commercial insurance. At 12 months, 23.7 percent remained on therapy. At 24 months, that figure dropped below 15 percent.
A 2023 JAMA Network Open study covering a broader commercial claims dataset confirmed the range: one-year persistence between 20 and 28 percent depending on the drug and payer, with semaglutide (Ozempic/Wegovy) at the higher end of that window and older agents considerably lower. The spread matters — the newest, most-effective formulations do modestly better at keeping patients — but no real-world dataset has found anything close to the trial completion rates, which typically exceeded 80 percent.
The obvious objection is that trial participants are different: screened, supported, compensated for their time, and systematically followed up. That is exactly the point. The 17-plus percent weight loss in STEP-1 was achieved in a setting no health system replicates. Applying that number to population-level expectations is not wrong because it is cherry-picked — the data is real. It is wrong because the population it describes is not the population taking the drug at scale.
Why patients stop — and what the data actually shows
The reasons for discontinuation cluster into three buckets, and the proportion of each changes dramatically depending on which study you read. In randomized trials, the primary discontinuation reason is side effects — nausea, vomiting, gastrointestinal events account for most of the roughly 7 to 10 percent who drop out. That is the number pharma communicates, and it is accurate within the trial context.
In real-world claims and survey data, the picture inverts. Cost and insurance access are the dominant reasons. A 2023 survey by the American Society for Metabolic and Bariatric Surgery found that among patients who had stopped a GLP-1, 47 percent cited out-of-pocket cost or insurance denial as the primary reason. Side effects, by comparison, drove about 19 percent of real-world discontinuations in the same survey.
The third bucket — patients who stopped because they achieved their goal and felt they did not need to continue — is smaller than intuition suggests. Multiple datasets put it at under 10 percent. Most people who stop are not stopping because it worked well enough. They are stopping because they cannot afford it, or because their payer decided they no longer qualified.
The 23 percent who stay are not proof that the drug works for some and not others. They are a selection effect produced by insurance coverage and ability to pay.
The re-initiation data nobody tracks properly
One finding that I think gets undercovered: real-world GLP-1 use is often discontinuous. The JAMA Network Open 2023 analysis found that among patients who stopped within 12 months, roughly 30 to 40 percent had at least one re-initiation episode within the same observation window. Some of these are patients cycling off due to side effects and restarting. Some are coverage gaps. Some are intentional drug holidays.
This creates an interpretation problem. A patient who is off-drug for 3 months and back on for 3 months in a rolling 12-month window is a "discontinuation" in most persistence analyses. Whether their long-term outcomes look more like the continuous-use group or the never-restarted group is an empirical question we do not have great data on yet. The SURMOUNT-5 trial design acknowledged this by including a drug-cycling arm — but it is not finished.
My read: intermittent use is almost certainly less effective than the trial protocol, but quite possibly more effective than the press narrative of "stop and gain it all back" suggests. That narrative describes abrupt, no-restart withdrawal. Most real-world cycling is messier and more varied than that.
What trial designs systematically exclude
The 68-week duration of STEP-1, STEP-3, and SURMOUNT-1 was not chosen at random. It is long enough to show meaningful, defensible weight-loss curves — and short enough that cost-related attrition, which accelerates sharply after the first refill authorization period, does not swamp the efficacy signal. This is not a conspiracy; it is how drug development works. You design a trial to demonstrate what the drug does when taken as directed. The question of whether people will take it as directed is left to post-marketing.
The post-marketing reality in the US, where Wegovy listed at around $1,300 per month and many insurers covered it inconsistently or not at all through 2023 and 2024, is the 23 percent figure. In countries with subsidized access — Denmark, where semaglutide originated, or the UK, where NICE approved it with NHS coverage — persistence rates are meaningfully higher. The drug is the same. The access structure is not.
The counter-argument: 23 percent still matters
I want to give the strongest version of the other side. Even at 23 percent 12-month persistence, that is roughly 7 million Americans on therapy at any given time based on current prescription volumes. For those patients — the ones with consistent coverage, manageable side effects, and the structural support to continue — the efficacy numbers from the trials are probably a reasonable guide to what they will achieve. The drug does not work worse for them because other people stopped.
The problem is not that the 17 percent weight loss figure is false. The problem is that every policy claim, coverage decision, and investor valuation built on "GLP-1 will solve obesity at scale" is implicitly assuming something closer to trial-level persistence — and that assumption is not justified by the available data. If real-world persistence stays at 20 to 25 percent, the drug class will produce real benefits for a subset of users while leaving the population-level obesity numbers largely unchanged. That is a smaller, more honest story than the one being told.
What this changes
In practical terms, the discontinuation data reframes three questions. First, it reframes what "access" means — getting a prescription is not the same as sustained use, and policies that increase initiation without addressing cost barriers will likely shift the persistence curve very little. Second, it reframes the economic case for employer and payer coverage — if 75 percent of initiators stop within a year, the cost-effectiveness calculus that assumes multi-year weight maintenance needs to be rerun with realistic adherence inputs. Third, it reframes the clinical conversation — the right question for a patient starting a GLP-1 is not just "what is the expected weight loss" but "what is the plan when coverage changes or costs rise, which is the more likely scenario."
None of this is an argument against the drug class. It is an argument for using the data that exists rather than the data that is easier to announce. The 23 percent figure is in the public literature. It did not make the press release. That is a choice worth naming.
Ozemback — June 2026